Update app.py

app.py

@@ -1,17 +1,21 @@
 import os
 import numpy as np
 import torch
+import torchaudio
 
 # --- CRITICAL ENVIRONMENT FIXES ---
 # 1. Fix for Hugging Face millicore OMP_NUM_THREADS error (e.g., "7500m")
-# This prevents the underlying math libraries from crashing on startup.
 if os.environ.get("OMP_NUM_THREADS", "").endswith("m"):
     os.environ["OMP_NUM_THREADS"] = "1"
 
-# 2. Monkeypatch torch.load to default weights_only=False
-# PyTorch 2.6 changed the default to True for security, which breaks 
-# WhisperX/Pyannote checkpoints. Monkeypatching is the only way to fix 
-# internal library calls that we don't control.
+# 2. Force Torchaudio Backend to avoid VAD hang
+try:
+    if "ffmpeg" in torchaudio.list_audio_backends():
+        torchaudio.set_audio_backend("ffmpeg")
+except Exception:
+    pass
+
+# 3. Monkeypatch torch.load to default weights_only=False for PyTorch 2.6+
 import torch.serialization
 original_load = torch.load
 def patched_load(*args, **kwargs):
@@ -48,7 +52,7 @@ try:
 except Exception as e:
     print(f"Safe Globals Warning: {e}")
 
-# Fix NumPy 2.0+ attribute removal (required for older pyannote internals)
+# Fix NumPy 2.0+ attribute removal
 if not hasattr(np, 'NaN'):
     np.NaN = np.nan
 
@@ -144,14 +148,24 @@ with st.sidebar:
     
     st.header("Model Settings")
     model_size = st.selectbox("Whisper Model", ["large-v2", "medium", "base"], index=0)
+    
+    # --- New Language Option ---
+    language_map = {
+        "Auto-Detect": None,
+        "English": "en",
+        "Spanish": "es",
+        "French": "fr",
+        "German": "de",
+        "Italian": "it",
+        "Portuguese": "pt"
+    }
+    selected_lang_label = st.selectbox("Audio Language (Speeds up processing)", list(language_map.keys()), index=1)
+    target_language = language_map[selected_lang_label]
+    
     num_speakers = st.number_input("Speakers (0=Auto)", min_value=0, value=0)
     
     st.divider()
     st.info("API Keys are managed via Environment Secrets.")
-    if not ACTIVE_GEMINI_KEY:
-        st.error("⚠️ Gemini API Key not found!")
-    if not ACTIVE_HF_TOKEN:
-        st.error("⚠️ HF Token not found!")
 
 uploaded_file = st.file_uploader("Upload Video/Audio Clip", type=["mp4", "m4a", "wav", "mp3", "mov"])
 
@@ -168,7 +182,7 @@ if uploaded_file:
                     with open("temp_input", "wb") as f:
                         f.write(uploaded_file.getbuffer())
                     
-                    st.write("🎵 **Extracting Audio...**")
+                    st.write("🎵 **Extracting Audio (WAV 16k Mono)...**")
                     subprocess.run([
                         "ffmpeg", "-i", "temp_input", 
                         "-vn", "-acodec", "pcm_s16le", "-ar", "16000", "-ac", "1", 
@@ -177,7 +191,10 @@ if uploaded_file:
                     
                     try:
                         device = "cuda" if torch.cuda.is_available() else "cpu"
-                        st.write(f"🚀 **Running WhisperX on {device}...**")
+                        if device == "cpu":
+                            st.warning("⚠️ No GPU detected. WhisperX will be very slow.")
+                            
+                        st.write(f"🚀 **Loading WhisperX on {device}...**")
                         
                         # 1. Transcribe
                         compute_type = "float16" if device == "cuda" else "int8"
@@ -185,9 +202,11 @@ if uploaded_file:
                         
                         st.write("📝 **Transcribing...**")
                         audio = whisperx.load_audio("temp_audio.wav")
-                        result = model.transcribe(audio, batch_size=16)
                         
-                        # Memory management
+                        # Pass the language to speed up processing
+                        result = model.transcribe(audio, batch_size=16, language=target_language)
+                        
+                        # Memory cleanup
                         del model
                         gc.collect()
                         torch.cuda.empty_cache()
@@ -208,7 +227,7 @@ if uploaded_file:
                         diarize_kwargs = {}
                         if num_speakers > 0:
                             diarize_kwargs = {"min_speakers": num_speakers, "max_speakers": num_speakers}
-                            
+                        
                         diarize_segments = diarize_model(audio, **diarize_kwargs)
                         
                         # 4. Final Merge